Skip to main content

kl_div_loss_forward

teeny_kernels::nn::loss::bce

Function kl_div_loss_forward 

Source 
pub fn kl_div_loss_forward<T: Triton, const BLOCK_SIZE: i32>(
    input_ptr: T::Pointer<f32>,
    target_ptr: T::Pointer<f32>,
    out_ptr: T::Pointer<f32>,
    n_elements: i32,
)
where
    T::I32Tensor: Tensor<i32, 1> + Comparison<i32, BoolTensor = T::BoolTensor>,
    T::Pointer<f32>: AddOffsets<i32, 1, T::I32Tensor, Output = T::Tensor<T::Pointer<f32>>>,
Expand description

Element-wise KL-divergence forward.

PyTorch convention: input is log-probability, target is probability.

out = target * (log(target) - input)

Masked: out = 0 where target <= 0.